Reproduce “SmoothGrad: removing noise by adding noise” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, Martin Wattenberg
Researchers at Google Inc. present SmoothGrad, a simple yet effective method to reduce visual noise in gradient-based sensitivity maps, which are used to interpret deep learning models. The approach averages gradients over multiple noisy input samples, producing more visually coherent and discriminative explanations of what a model focuses on. Applying SmoothGrad improved the interpretability of maps from models like Inception v3 and showed additive benefits when combined with training networks using noise.
Key insights
- The visual noisiness of gradient-based sensitivity maps stems from sharp, local fluctuations in the derivative of the class score function.
- Averaging gradients over a local neighborhood in the input space can effectively smooth out these erratic fluctuations, yielding more coherent sensitivity maps.
- Training neural networks with noise also leads to smoother sensitivity maps, and this effect is complementary to applying SmoothGrad during inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.